CN107977619B - Hyperspectral target detection method based on ensemble learning constraint energy minimization - Google Patents

Abstract

The invention relates to a hyperspectral target detection method based on ensemble learning constrained energy minimization, which comprises the following five steps: the method comprises the following steps: reading data by a computer; step two: expanding a data set; step three: randomly sampling to obtain a plurality of detectors; step four: calculating an included angle between the pixel vector and the interested target; step five: and performing soft voting on the obtained distribution condition. The method overcomes the defects of the prior art, well solves the problem of hyperspectral image target detection, and obtains a better detection result, so that the method can be applied to hyperspectral image target detection and has wide application prospect and value.

Description

Hyperspectral target detection method based on ensemble learning constraint energy minimization

The technical field is as follows:

the invention relates to a hyperspectral target detection method based on ensemble learning constrained energy minimization, and belongs to the technical field of hyperspectral remote sensing image target detection.

(II) background technology:

the hyperspectral remote sensing image not only contains the spatial information of an object, but also contains the spectral information of the object, wherein each pixel contains an approximately continuous spectral curve. All pixels in the hyperspectral image in the same wave band can form a two-dimensional image, and each pixel also comprises a spectrum curve, so that the hyperspectral image data can be regarded as three-dimensional data and contain abundant spectrum information. The hyperspectral image has high spectral resolution, can distinguish materials of different ground objects through spectral characteristics, and can solve the problems which cannot be solved by using the original full-color image and multispectral image, such as military camouflage identification, underground work, resource detection, environment monitoring and the like, so that the hyperspectral image target detection has important application value in military and civil aspects.

The existing algorithms for detecting the hyperspectral image target include a Spectral Angle Matching (SAM) algorithm, a Spectral Information Divergence (SID) algorithm, a Constrained Energy Minimization (CEM) algorithm, a Matched Filter (MF), and an adaptive coherence/cosine estimator (ACE).

Most detectors constructed by the existing algorithm are obtained by a constraint least square method or a hypothesis testing algorithm based on Gaussian distribution prior, and most of the obtained detectors are linear operators (such as CEM/MF) or quadratic operators (ACE). However, the hyperspectral data captured in the real environment is usually affected by imaging noise, atmospheric turbulence, spectrum mixing and other effects, usually shows strong nonlinear distribution characteristics, and does not satisfy gaussian distribution prior in the spectrum space. In this case, it is often difficult to obtain satisfactory detection results by linear operators or quadratic operators. Although some scholars in the field of hyperspectral target detection map hyperspectral data points to a high-dimensional or infinite-dimensional space through a kernel method (kernel trim), nonlinear expression forms of the above algorithms such as kernel cem (kernel cem), kernel ace (kernel ace) and the like are obtained, the kernel method itself is easily subjected to kernel function parameter selection and shows a phenomenon of unstable performance. From the theory of statistical machine learning, the expression ability and generalization ability of the algorithm are two important factors restricting the accuracy and stability of the algorithm, for traditional algorithms such as CEM, MF and ACE, the generalization ability is strong, the expression ability is relatively weak, the spectrum data under nonlinear distribution is difficult to describe, and the nucleation forms of the traditional algorithms have strong expression ability and weak generalization performance. In general, the key to designing a hyperspectral target algorithm is the trade-off between algorithm expressive power and generalization power.

An integrated learning algorithm is an important method in machine learning, and the idea is to improve the expression capacity and generalization capacity of the integrated learning algorithm by adopting the integration of a plurality of weak decision makers. Two important ideas exist in ensemble learning, one is called Boosting, the effect of Boosting is to keep the generalization performance of a weak decision maker unchanged and simultaneously improve the expression capacity of the weak decision maker, and a typical algorithm is an Adaboost algorithm; and the other type of idea called Model Combination is to ensure the strong expression capability of the weak decision maker and simultaneously improve the generalization performance of the weak decision maker, and the typical algorithm is Bagging. Therefore, one problem that we can easily think is that the existing hyperspectral target detection algorithm can be improved by using the idea of Model Combination in ensemble learning, so that the hyperspectral target detection algorithm has strong nonlinear spectral expression capability and improves the stability and robustness of the hyperspectral target detection algorithm. Aiming at the situation, the invention provides an Ensemble-based-CEM (E-CEM for short) hyperspectral target detection method with minimized constraint energy by Ensemble learning.

(III) the invention content:

1. the purpose is as follows: the invention aims to provide a hyperspectral target detection method for minimizing the constraint energy of ensemble learning.

2. The technical scheme is as follows: the invention is realized by the following technical scheme:

the invention discloses a hyperspectral target detection method based on ensemble learning constraint energy minimization. The method comprises the following specific steps:

the method comprises the following steps: the computer reads the data.

Step two: expanding the data set, and specifically adopting a QCEM algorithm;

step three: extracting N samples from the data set, repeating the step two P times to obtain M sampling sets containing N training samples, training each sampling set to obtain a CEM detector, and obtaining the distribution condition of the interested target d by using the detectors;

step four: and calculating the included angle between each pixel vector in the data set and the interested target.

Step five: and performing soft voting on the distribution condition of the interested target d in the data set, which is obtained by all the CEM operators, and finally obtaining the distribution condition of the interested target d in the whole data set by taking the included angle obtained in the step four as the weight.

3. The advantages and the effects are as follows:

the invention discloses a hyperspectral target detection method based on ensemble learning constraint energy minimization. The invention has the advantages that: the hyperspectral target detection is completed by a method of integrating learning and constraining energy minimization. The stability and robustness of the hyperspectral target are integrally improved on the basis that the constraint energy minimization has stronger spectral expression capability, and the overall effect of hyperspectral target detection is better through the combination of the constraint energy minimization and the hyperspectral target.

(IV) description of the drawings:

FIG. 1 is a block diagram of hyperspectral target detection by the method of the invention.

FIG. 2 is the AUC value of the method of the present invention and other related algorithms on hyperspectral data.

(V) specific embodiment:

for a better understanding of the technical solution of the present invention, the following embodiments of the present invention are further described with reference to the accompanying drawings:

the invention is implemented in the MATLAB 2016a programming environment. After reading the hyperspectral remote sensing image data, the computer firstly expands the data set, then replaces and extracts N sample pixels to train a CEM detector, repeats T times to obtain T detectors, judges the distribution condition of an interested target on the data set by using the trained detectors, performs soft voting by taking the included angle of a pixel vector and a target vector as a weight to obtain the final distribution condition of the interested target in the data set, and finally completes hyperspectral target detection.

The flow chart of the invention is shown in 1, a computer is configured by adopting an Intel (R) core (TM) i7-3770 processor, the main frequency is 3.40GHz and the memory is 32GB, and the hyperspectral target detection method comprises the following steps:

the method comprises the following steps: the computer reads the data. The method comprises the steps of reading an AVIRIS hyperspectral image by using a computer in an MATLAB 2016a programming environment, wherein data used by the method is an Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) hyperspectral image for target detection, and the scene of the image is an airport in San Diego in the United states. The data collected by the AVIRIS have 224 wave bands, and the wave band range is 0.4-2.5 μm. The water vapor absorption band and the low signal-to-noise ratio band are removed first, the actually used data have 189 bands, and the image size is 200 x 200 pixels. Recording a matrix S ═ r formed by all pixels in the hyperspectral image₁，r₂，...，r_NIn which r is_j＝(r_j1，r_j2，...，r_jL)^TThe image is any sample pixel vector (j ═ 1, 2.., N), N is the number of pixels, and L is the number of bands of the image. The object of interest is d ═ d (d)₁，d₂，...，d_L)。

Step two: and expanding the data set. Adopting QCEM algorithm;

to achieve a stronger nonlinear representation capability than the Constrained Energy Minimization (CEM) algorithm, we here use the modified algorithm QCEM. The QCEM algorithm is to design the filter G ═ diag (G)₁，g₂，...，g_L) And ω ═ ω (ω)₁，ω₂，...，ω_L)^T. Where the filter can be seen as a matrix, g₁，g₂，...，g_LAnd ω₁，ω₂，...，ω_LRespectively, represents an element in a matrix, and T represents a matrix transpose.

When the input is r_jWhen it is, remember y_jFor the output through the filter:

wherein, T represents the matrix transposition,

is r_jThe transposed matrix of (2).

G, ω can be solved by:

wherein, E { y₂The desired value of output energy is shown as beta (/)_2F+∥G∥_2F) Is a regularization term to reduce the complexity and improve the stability of the filter.

The pixel vectors in the dataset are expanded as:

the object of interest is expanded into

The number of bands is expanded to 2L.

(2) The formula can be converted into:

obtaining:

step three: extracting N samples from the data set, repeating the step two times and P times to obtain M sampling sets { D) containing N training samples₁，D₂，...，D_PTherein of

For any one of the sample sets, the sampling rate is,

for any pixel vector (i ═ 1, 2.., N) in the sample set, each sample set D_iA CEM operator can be obtained, the CEM operator is acted on each pixel in the data set, and the distribution condition of the interested target d in the data set is obtained

Wherein

Is h_iIn that

The final distribution of the objects of interest obtained by all CEM operators can be expressed as { c }₁，c₂，...，c_NTherein of

Describing an algorithm:

inputting: data set S ═ r₁，r₂，...，r_NWhere r_i＝(r_i1，r_i2，...，r_iL)^T；

A target spectrum of interest d;

extracting N pixel vectors M in a replacement mode;

CEM algorithm

The number of training rounds P;

the process is as follows:

step four: calculating the included angle (theta) between each pixel vector in the data set and the interested target d₁，θ₂，...，θ_NLet α_i＝1-θ_iTo obtain { alpha₁，α₂，...，α_N}。α_iThe calculated value of the included angle theta changes along with the change of the similarity degree of the pixel and the interested target, if the pixel is more similar to the interested target, the included angle of the two vectors is smaller, and alpha is_iThe larger, conversely, the larger alpha_iThe smaller, { α₁，α₂，...，α_NCan be used as the weight in soft voting.

Step five: distribution of objects of interest in the dataset by all CEM operators { c }₁，c₂，...，c_NPerforming soft voting, and obtaining an included angle { alpha ] in the step four₁，α₂，...，α_NAnd f, taking the weight as the weight, and finally obtaining the distribution condition of the interested target d in the whole data set.

The experimental results are as follows: in order to verify the effectiveness of the method, the method is used for carrying out experiments, and a good hyperspectral target detection effect is obtained. The data used in the experiment of the invention is an Airborn Visible/Infrared Imaging Spectrometer (AVIRIS) hyperspectral image. The Area Under the ROC Curve (AUC) of the hyperspectral target detection method (E-CEM) with ensemble learning constrained energy minimization finally obtained on test data is 0.9982, and the hyperspectral target detection effect is good (see fig. 2 and table 1 below).

TABLE 1

From the experimental result, the method well solves the problem of visible hyperspectral target detection, so that the method can be applied to hyperspectral target detection and has wide application prospect and value.

Claims (1)

1. A hyperspectral target detection method based on ensemble learning constraint energy minimization is characterized by comprising the following steps: the method comprises the following steps:

the method comprises the following steps: reading data by a computer; reading the AVIRIS hyperspectral image by using a computer in an MATLAB 2016a programming environment, and performing target detection on the AVIRIS hyperspectral image by using data, wherein the scene of the image is an airport in san Diego in the United states; the data collected by the AVIRIS have 224 wave bands, and the wave band range is 0.4-2.5 mu m; firstly, removing water vapor absorption wave bands and low signal-to-noise ratio wave bands, wherein actually used data have 189 wave bands, and the size of an image is 200 x 200 pixels; recording a matrix S ═ r formed by all pixels in the hyperspectral image₁，r₂，...，r_NIn which r is_j＝(r_j1，r_j2，...，r_jL)^TThe method comprises the steps that any sample pixel vector is obtained, j is 1, 2, the. The object of interest is d ═ d (d)₁，d₂，...，d_L)；

Step two: expanding the data set; adopting QCEM algorithm;

in order to obtain the nonlinear expression capability stronger than that of the constrained energy minimization CEM algorithm, an improved algorithm QCEM is adopted; the QCEM algorithm is to design the filter G ═ diag (G)₁，g₂，...，g_L) And ω ═ ω (ω)₁，ω₂，...，ω_L)^T(ii) a Where the filter is considered as a matrix, g₁，g₂，...，g_LAnd ω₁，ω₂，...，ω_LRespectively representing elements in the matrix, and T represents matrix transposition;

when the input is r_jWhen it is, remember y_jFor the output through the filter:

wherein, T represents the matrix transposition,

is r_jThe transposed matrix of (2);

g, ω is solved by:

wherein, E { y₂The expected value of output energy is shown, and beta (//w//2F +/G// 2F) is a regularization term used for reducing the complexity of the filter and improving the stability of the filter;

the pixel vectors in the dataset are expanded as:

the object of interest is expanded into

The number of wave bands is expanded to 2L;

(2) the formula is converted into:

obtaining:

step three: extracting N samples from the data set, repeating the step two times and P times to obtain M sampling sets { D) containing N training samples₁，D₂，...，D_PTherein of

For any one of the sample sets, the sampling rate is,

for any pixel vector in the sampling set, i is 1, 2_iObtaining a CEM operator, and applying the CEM operator to each pixel in the data set to obtain the distribution condition of the target d of interest in the data set

Wherein

Is h_iIn that

The final distribution of the objects of interest obtained by all CEM operators is expressed as { c }₁，c₂，...，c_NTherein of

Describing an algorithm:

inputting: data set S ═ r₁，r₂，...，r_NWhere r_i＝(r_i1，r_i2，...，r_iL)^T；

A target spectrum of interest d;

extracting N pixel vectors M in a replacement mode;

CEM algorithm

The number of training rounds P;

the process is as follows:

1：

2：for i＝1，2，...，P do3：

4：

5：end for

step four: calculating the included angle (theta) between each pixel vector in the data set and the interested target d₁，θ₂，...，θ_NLet α_i＝1-θ_iTo obtain { alpha₁，α₂，...，α_N}；α_iThe calculated value of the included angle theta changes along with the change of the similarity degree of the pixel and the interested target, if the pixel is more similar to the interested target, the included angle of the two vectors is smaller, and alpha is_iThe larger, conversely, the larger alpha_iThe smaller, { α₁，α₂，...，α_NTaking the weight as the weight of soft voting;

step five: distribution of objects of interest in the dataset by all CEM operators { c }₁，c₂，...，c_NPerforming soft voting, and obtaining an included angle { alpha ] in the step four₁，α₂，...，α_NAnd f, taking the weight as the weight, and finally obtaining the distribution condition of the interested target d in the whole data set.

Images